Multi-Agent Orchestration Troubleshooting

Solve common issues with Agentforce solutions involving Multi-agent Orchestration.

How can I reduce latency in my Multi-agent Orchestration solution?

Start by implementing the recommendations in "Ways to Minimize Latency in Multi-Agent Solutions" in Multi-Agent Orchestration.
Isolate where high latency and timeouts occur in your Agentforce solution. Use debugging tools described in "Troubleshooting Resources." For example:
- In Agent Analytics, monitor "duration" metrics and scores for your solution. In Agent Health Monitoring, set alert thresholds for relevant metrics and scores to trigger notifications when latency breaches threshold levels.
- Investigate traces and logs for timeout errors. The timeout limit for orchestrator agents is 120 seconds, which includes the elapsed time spent by the orchestrator agent doing overhead tasks, such as delegating and (in supervisor mode) synthesizing the results. In supervisor mode, the orchestrator agent synthesizes timeout errors. The timeout limit for connected subagents is 30 seconds.
Use these symptoms to uncover latency root causes that you can remedy in your AI solution.

Keep in mind that:

Custom (mutable) variables can be specified during the session.
Context (linked) variables are immutable and can’t be updated after the session has started.

In Multi-agent Orchestration, there are two ways in which to pass variables:

From an orchestrator agent to a connected subagent. For connected subagents, input variables can be mapped. In both handoff and supervisor mode, you can map these input variables to either context or custom variables. See "Handoff and Supervisor Modes" in Multi-Agent Orchestration.
Note In this release, variables are passed in one direction only: from the orchestrator agent to the connected subagent. Variables aren’t passed from a connected subagent back to the orchestrator agent.
Between connected subagents. In handoff mode, if a connected subagent has an After Response section, you can set variables before transitioning to another connected subagent. Because the next connected subagent receives these values only through its inputs block, any variables you set require mapping in the inputs block of the target connected subagent. Context doesn’t flow automatically.

See Connect an Agent as a Subagent in Salesforce Help and Connected Subagent Block in the Agentforce Developer Guide.

Your authentication approach varies by agent type and by the combination of orchestrator agent and connected subagent.

Agentforce Service Agent (ASA) connected subagents run as the orchestrator agent's session user (the digital user), not as its own agent user. ASA also supports unauthenticated guest users.
Agentforce Employee Agent (AEA) connected subagents run in the logged-in user's context. Assign the logged-in user the permissions required for any operation the connected subagent requests.

See "User-Level Authentication in Multi-Agent Solutions" in Multi-Agent Orchestration.

Review supported and unsupported functionality in Considerations and Limitations.
Multi-agent Orchestration enforces no limit on the number of connected subagents. Experiment with different combinations to come up with the optimal design that scales for your use case and environment.
For design recommendations, see "Tips for Designing Your Agent-to-Agent Solution" in Multi-Agent Orchestration.

Agentforce provides tools and features that help you identify patterns, investigate root causes, and take corrective action.

In Agent Builder, use the Plan Tracer in the Conversation Preview Panel to investigate issues in your multi-agent solution. Inspect the detailed, step-by-step decisions and actions your orchestrator agent and connected subagents take during a conversation or workflow test.
Use enhanced event logs to provide granular details about agent conversations.
Use Agent Analytics to monitor agent performance and detect error trends. Investigate patterns and root causes at the session level (see Analyze Data with Agent Optimization).
In Agent Health Monitoring, set alert thresholds to trigger notifications when agent performance breaches threshold levels.

此文章是否解決您的問題？

請讓我們知道，以便我們改進！